cadrumo.application.modelo.pulled_filing_reconcile module¶

Reconcile a local calculation against this taxpayer’s own pulled AEAT filing.

The filed-history sweep persists, per pullable modelo, the per-casilla values AEAT holds for a filing the taxpayer already made. The local calculation for the same modelo and period persists its own casilla values. Both sides live in the same encrypted bucket, and nothing compared them: a taxpayer whose local figures disagree with what AEAT records them as having filed learned it from neither surface.

WHAT THIS IS NOT. It is not the engine-drift check. That already ships: the filed-state verification path re-runs the registry calculation using the FILED declaration’s own input casillas and compares the computed outputs, which asks whether this engine reproduces AEAT’s arithmetic from AEAT’s inputs. It never consults the taxpayer’s own calculation, so it cannot answer whether the taxpayer’s working disagrees with their filing. That second question is this module’s, and the two must not be collapsed.

WHY IT IS SCOPED RATHER THAN TOTAL. A freshly onboarded profile computes zero nearly everywhere, so comparing every pulled figure against it would raise a mismatch on essentially every reconciled casilla — the alert fatigue that trains an operator to ignore a channel. The comparison is therefore restricted to the casillas the local revision supplied independent evidence for, resolved by resolve_casilla_population_scope(). An untouched bucket produces no findings at all, which is the intended behaviour and not a failure to detect.

GROUNDING IS CARRIED, NEVER MINTED. A divergence finding makes no independent legal claim: it says the value the registry grounds at these references does not match what was filed. So each finding carries the diverging casilla’s own legal_refs and source_refs rather than references invented for the finding. The registry schema types both as minimum-length-one tuples on every casilla definition, so there is always something to carry and no substitution path exists.

THE AGREEMENT THRESHOLD IS THE REGISTRY’S, NOT THIS MODULE’S. How close two values must be to count as agreeing is a regulatory question versioned by filing year and revision: the registry publishes a tolerance per verification expectation and folds the strictest across them. Some revisions publish one cent; others publish 0.00 and so demand exact equality. This module therefore reads verification_policy().tolerance from the law-resolved snapshot rather than carrying a constant of its own — a hardcoded cent would silently absorb a one-cent divergence on every revision of the second kind, which is an under-declaration this channel exists to surface.